Edinburgh Research Explorer

Reducing The Computational Requirements for Simulating

Tunnel Fires by Combining Multiscale Modelling and Multiple
Processor Calculation

Citation for published version:

Vermesi, I, Rein, G, Colella, F, Valkvist, M & Jomaas, G 2018, 'Reducing The Computational Requirements
for Simulating Tunnel Fires by Combining Multiscale Modelling and Multiple Processor Calculation',
Tunnelling and Underground Space Technology. https://doi.org/10.1016/j.tust.2016.12.016

Digital Object Identifier (DOI):

10.1016/j.tust.2016.12.016

Link:
Link to publication record in Edinburgh Research Explorer

Document Version:
Publisher's PDF, also known as Version of record

Published In:
Tunnelling and Underground Space Technology

General rights
Copyright for the publications made accessible via the Edinburgh Research Explorer is retained by the author(s)
and / or other copyright owners and it is a condition of accessing these publications that users recognise and
abide by the legal requirements associated with these rights.

Take down policy

The University of Edinburgh has made every reasonable effort to ensure that Edinburgh Research Explorer
content complies with UK legislation. If you believe that the public display of this file breaches copyright please
contact openaccess@ed.ac.uk providing details, and we will remove access to the work immediately and
investigate your claim.

Download date: 11. Apr. 2020

 Tunnelling and Underground Space Technology 64 (2017) 146–153

Contents lists available at ScienceDirect

Tunnelling and Underground Space Technology

journal homepage: www.elsevier.com/locate/tust

Reducing the computational requirements for simulating tunnel fires by

combining multiscale modelling and multiple processor calculation
Izabella Vermesi a,⇑, Guillermo Rein b, Francesco Colella c, Morten Valkvist d, Grunde Jomaas a
a
Department of Civil Engineering, Technical University of Denmark, 2800 Kongens Lyngby, Denmark
b
Department of Mechanical Engineering, Imperial College London, SW7 2AZ London, UK
c
Exponent Inc., 9 Strathmore Rd, Natick, MA, USA
d
Greater Copenhagen Fire Department, Bag Rådhuset 3, 1550 Copenhagen W, Denmark

a r t i c l e i n f o a b s t r a c t

Article history: Multiscale modelling of tunnel fires that uses a coupled 3D (fire area) and 1D (the rest of the tunnel)
Received 9 December 2015 model is seen as the solution to the numerical problem of the large domains associated with long tunnels.
Received in revised form 4 December 2016 The present study demonstrates the feasibility of the implementation of this method in FDS version 6.0, a
Accepted 30 December 2016
widely used fire-specific, open source CFD software. Furthermore, it compares the reduction in simula-
Available online 15 February 2017
tion time given by multiscale modelling with the one given by the use of multiple processor calculation.
This was done using a 1200 m long tunnel with a rectangular cross-section as a demonstration case. The
Keywords:
multiscale implementation consisted of placing a 30 MW fire in the centre of a 400 m long 3D domain,
CFD
Multiscale modelling
along with two 400 m long 1D ducts on each side of it, that were again bounded by two nodes each. A
Tunnel fires fixed volume flow was defined in the upstream duct and the two models were coupled directly. The fea-
FDS6 sibility analysis showed a difference of only 2% in temperature results from the published reference work
Computational efficiency that was performed with Ansys Fluent (Colella et al., 2010). The reduction in simulation time was signif-
icantly larger when using multiscale modelling than when performing multiple processor calculation
(97% faster when using a single mesh and multiscale modelling; only 46% faster when using the full tun-
nel and multiple meshes). In summary, it was found that multiscale modelling with FDS v.6.0 is feasible,
and the combination of multiple meshes and multiscale modelling was established as the most efficient
method for reduction of the calculation times while still maintaining accurate results. Still, some unphys-
ical flow oscillations were predicted by FDS v.6.0 and such results must be treated carefully.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction (Migoya et al., 2009). As an alternative to sacrificing either com-

Fire modelling is frequently used as a means of investigating a
variety of fire scenarios. Whereas such modelling is feasible for
most type of structures, long tunnels are problematic from a
numerical point of view, because they, due to their length, are
defined by large domains that require very large computational
resources. This numerical challenge is amplified by the fact that
finding the proper fire safety strategy often requires trying a num-
ber of scenarios to establish all the essential characteristics of the
system. Among the models found in literature, Vega et al. (2008)
required 50 h to simulate 10 min of a fire using a 3D model in
ANSYS Fluent. Other models use either short tunnels (Jain et al.,
2008) or simplified one-dimensional models for longer tunnels
⇑ Corresponding author at: Department of Mechanical Engineering, Imperial
College London, UK.
E-mail address: i.vermesi14@imperial.ac.uk (I. Vermesi).
plexity or time efficiency, multiscale modelling for tunnel flows
and fires has previously been studied using a general purpose com-
putational fluid dynamics (CFD) software and it has yielded satis-
factory results in comparison to full scale CFD simulations
(Colella et al., 2010). The method combines a 3D domain for the
near fire zones, which are characterized by large temperature
and pressure gradients, with a 1D network approach for the bulk
flow in the far field. A full description of the concept, as well as
an assessment of the method’s accuracy, is presented by Colella
et al. (2010, 2011). The model used in Colella’s work has been val-
idated using experimental measurements from real tunnel flows
(Colella et al., 2010). As the previous study used ANSYS Fluent,
there is a need for a feasibility study using other modelling
techniques.
Herein, the primary investigation is thus the feasibility of mul-
tiscale modelling of tunnel fires in Fire Dynamics Simulator (FDS
v.6.0), which is a CFD model for fire driven flow that is widely used

http://dx.doi.org/10.1016/j.tust.2016.12.016
0886-7798/Ó 2017 Elsevier Ltd. All rights reserved.
 I. Vermesi et al. / Tunnelling and Underground Space Technology 64 (2017) 146–153
for heat and smoke transport from fires (Jain et al., 2008). It is
worth noting the difference in the governing equations in ANSYS
Fluent and FDS: the former uses Reynolds-averaged Navier Stokes
equations (RANS) for its simulations, whereas the latter uses Large
Eddy Simulations. A comprehensive analysis of the differences
between RANS and LES modelling are found in the work by
Versteeg and Malalasekera (2007). Furthermore, FDS is an open-
source software, thus it is very widely used, especially in the indus-
try. Therefore, developing a feasible multiscale model for this soft-
ware can make a significant impact on reducing simulation times
for a large number of users. Parametric studies are needed in order
to obtain the relevant scenario that is going to be analyzed in detail
for the final tunnel ventilation design. This is possibly performed
with more computationally heavy programs or settings. As such,
the fast computation enables the designer to eliminate several sce-
narios in the process towards the final design.
The implementation of the multiscale model in FDS v.6.0,
sketched in Fig. 1, followed the work by Colella et al. (2010), while
taking advantage of the fire-specific capabilities of FDS v.6.0. The
geometry and model guidelines such as the domain length and vol-
umetric flow induced by the ventilation system served as a start for
the present model. The 3D component was created using the tradi-
tional 3D grid, while the 1D network was implemented using the
Heating, Ventilation and Air Conditioning (HVAC) feature of FDS
v.6.0, which is described in detail in the next section. These two
models are coupled directly in the FDS code, something which pro-
vides a continuous interaction between them. This is in contrast
with an indirect coupling that requires the use of an additional
software to model the 1D network, whose results are then used
as input for the boundary conditions of the 3D model. As a result
of this direct coupling, the time spent preparing the model is
reduced. The coupling of 3D and HVAC component in FDS has been
validated against real tunnel flow data in Ang et al. (2016), thus
confirming the possibility of coupling the two domains.
The second objective of this paper is to assess which of the fol-
lowing methods is the most time-efficient: multiscale modelling,
multiple processor calculations or a combination of the two.
Because of the computational requirements, the 3D grid was
divided in multiple meshes assigned to individual cores of the
computer. In this way, the order of magnitude of the duration of
the simulation decreased from weeks to hours. To find out which
method had contributed the most to the decrease in runtime dura-
tion, simulations were performed on the same model, first using a
single mesh in the 3D domain of the multiscale model to compare
with the multiscale-multiple-mesh model. Then the full tunnel
was simulated with a single mesh and multiple meshes.
2. Methodology
2.1. 3D model and fire scenario
In order to be able to compare the results of the method imple-
mented in FDS with the results obtained in the reference work
(Colella et al., 2010), the tunnel chosen for the analysis has a total
length of 1200 m and a longitudinal ventilation system. It is con-
1D network: 1 duct + 2
nodes (400m)
Direction of jet fan
interface between the models
Fig. 1. The concept of multiscale modelling of tunnels fires in FDS v.6.0.
147
sidered to be a road tunnel with traffic going in one direction only.
In the previous work, the tunnel cross-section had a horseshoe
shape with a height of 6.5 m and a cross-sectional area of 53 m2.
However, FDS has some constraints regarding geometry and can
only contain models with rectangular grids. An attempt was made
to simulate a circular cross section using a stair-stepped boundary
condition. However this is not a correct solution as it offers a differ-
ent behaviour of the flow, which is not necessarily a realistic beha-
viour. Therefore, an equivalent, rectangular cross-section was used
in the current study. In order to obtain an area similar to the pre-
vious one, the width was chosen to be 8 m, giving a cross-sectional
area of 52 m2. The hydraulic diameter is found to be
DH ¼ 4A P
¼ 7:17 m. The walls, floor and surface were defined as adi-
abatic concrete surfaces and the inlet and outlet connected to the
HVAC solver were defined as HVAC surfaces in order to permit
the interaction between the two models. The tunnel walls were
assumed as adiabatic for the sake of simplicity. Other heat transfer
boundary conditions to the walls could have been used, but the
adiabatic condition gives conservative estimates of the pressure
losses (highest fire throttling effect), back-layering velocity and
back-layering distance (Colella et al., 2010). Furthermore, the miss-
ing heat loss in the 1D part of the tunnel results in increased tem-
peratures and hence pressure losses, both of which in turn produce
a conservative estimate of the backlayering distance due to a
reduction in the airflow. Given that the main objective of the study
is to evaluate the performance of multiscale modelling in FDS v.6.0
compared to another CFD code, it is deemed acceptable to use adi-
abatic conditions.
In order for the simulation to give accurate results, the bound-
ary interface between the two models has to be placed at a location
where the flow is fully developed and the temperature or velocity
gradients are insignificant (Colella et al., 2012). As shown by
Colella et al. (2011), accurate results are obtained when the dis-
tance from the fire is at least 13 times larger than the hydraulic
diameter. Thus, a domain larger than 200 m should yield satisfac-
tory results. Results recorded 10 m from the fire become boundary
independent for grids larger than 200 m, whereas results recorded
100 m from the fire do not depend on the boundary interface for
grids larger than 400 m (Colella et al., 2010). Therefore, as a start-
ing point, the CFD domain was chosen to be 400 m along the lon-
gitudinal axis and the fire was placed in the centre, as is shown in
Fig. 2.
The fire used herein has a maximum heat release rate (HRR) of
30 MW, which can represent the peak heat release rate for a burn-
ing bus (Carvel and Beard, 2005). The fire is represented using a
lumped species approach for species tracking combined with a
mixing model and fast chemistry. The radiative fraction, which is
the amount of energy released from the fire as thermal radiation,
was chosen as the default radiative fraction in FDS, namely 0.35.
After an initial study of the fire development and change in mass
flow rate with respect to time (Colella et al., 2011), the simulation
duration was chosen as 600 s, at which point it was verified that
steady-state conditions had been reached throughout the tunnel.
1D network: 1 duct + 2
nodes (400m)
3D grid
(400m) interface with the ambient
148 I. Vermesi et al. / Tunnelling and Underground Space Technology 64 (2017) 146–153
H=6.5m
Length=400m
cross sectional
A=52sqm
Fig. 2. The 3D model representation of the tunnel (not to scale): cross-section (left) and longitudinal view (right).
2.1.1. Grid sensitivity analysis
As seen in previous studies, the results given by FDS are highly
dependent on the grid size (Petterson, 2002). Reducing the grid
size does not automatically mean a significantly better precision,
but it does considerably increase the runtime of the simulation.
Therefore, it is important to find a balance between the desired
precision and keeping the simulation time at a level acceptable
to the user.
It is important to state that the grid cells in FDS are recom-
mended to have a cubic shape (Floyd et al., 2013). This is possible
in the current case, thus each cell has the same width, length and
height. According to the User’s Guide for FDS (Floyd et al., 2013), an
initial sizing of the mesh cells should be done by evaluating the
non-dimensional expression D =@x where @x is the cell size and
D is the characteristic fire diameter. For the desired HRR of
30 MW, D ¼ 3:7 in this case, cell sizes should fall between
0.23 m and 0.92 m, according to results from previous studies
(Floyd et al., 2013). Therefore, a cell size of 0.40 m is chosen as
an initial setting. As the range above is just a guideline, a mesh sen-
sitivity analysis was performed for mesh sizes of 0.20 m, 0.25 m
and 0.80 m.
The results for temperature and longitudinal velocity, further
called u-velocity, were recorded at 210 m, which represent a dis-
tance of 10 m from the centre of the fire source. Recording devices
were placed every 0.5 m from a height of 0.5 m until 6 m in order
to obtain a more complete overview of the fire behaviour. Fig. 3
shows the average results over time for temperature and velocity
for each of the cell sizes used. 10 m away from the fire, the grid
using 0.80 m cells gives higher results for the temperature values,
but the 0.40 m, 0.25 m and 0.20 m cell grid show similar results.
The difference between the 0.40 m and the 0.25 m cell grid is not
significant, meaning that decreasing the cell size does not neces-
sarily induce a considerable increase in the precision. Also, the
results from the coarse mesh are quite stable compared with the
results from the moderate and finer meshes, which have some
oscillations in the mass flow due to a numerical issue in the solver.
The velocity results at both distances follow similar trends. The
coarse mesh predicts lower values of the flow velocity, whereas
Fig. 3. Grid sensitivity analysis: the average temperature (a) and the average u-velocity (b) results for 4 different cell size configurations.
the other ones are similar. The average results are presented in
Table 1, along with the variation from the results of the 0.20 m
mesh. The biggest difference is, as expected, in the mesh with cell
size of 0.80 m. The 0.40 m and 0.25 m meshes do not differ much
from the finest mesh, with none of the average values having dif-
ferences of more than 6%.
Therefore, in the case of temperatures, the errors are negligible,
with a variation of only a few degrees that would not influence the
overall fire safety design. As for the velocity, the average values for
the 0.20 m, 0.25 m and 0.40 m meshes are almost identical at
210 m and very similar at 300 m.
The above results show that there is no significant precision
improvement in decreasing the cell size from 0.40 m. The differ-
ences are negligible for the performance of the fire safety design.
However, the differences in runtimes are very significant. As
shown in Table 1, there is a very important increase in computa-
tional demand by decreasing the cell size. The difference in the
number of cells is reflected in the number of hours required for
the calculations. Calculations with the finest mesh take 92 h,
whereas calculations with the 0.40 m mesh are about 15 times fas-
ter. As the difference in precision does not justify using the finest
mesh, the 0.40 m cell mesh was chosen for the remainder of the
study.
2.1.2. Multiple meshes
FDS uses MPI (Message-Passing Interface) to allow multiple
computers to run a single FDS simulation (Floyd et al., 2013). The
FDS domain is divided into multiple meshes with equal cell sizes
that are then computed as different processes. This approach has
previously been used to produce satisfactory results for tunnel fire
simulations (Weisenpacher et al., 2011).
To ensure that the mesh division did not add significant errors
to the results of the simulation, a comparison was done using a sin-
gle mesh and 3 meshes, respectively. The domain was divided into
3 separate meshes, which were then assigned to individual pro-
cesses. The measurements were made 0.40 m away from the
boundary interface using temperature and velocity slice files.
Fig. 4 shows the velocity and the temperature profiles versus tun-
 I. Vermesi et al. / Tunnelling and Underground Space Technology 64 (2017) 146–153 149

Table 1
Comparison of the average results from simulations with different cell sizes, including runtime.

0.20 m 0.25 m 0.40 m 0.80 m

Average temperature at 210 m [°C] 126 128 124 152
Deviation – +2% 1% +21%
Average temperature at 300 m [°C] 114 118 112 110
Deviation – +4% 2% 3%
Average velocity at 210 m [m/s] 6.6 6.7 6.6 5.9
Deviation – +2% +1% 10%
Average velocity at 300 m [m/s] 7.4 7.3 7.0 6.5
Deviation – 0% 6% 12%
Runtime [h] 92 39 6 0.8

Fig. 4. Comparison of results for the 1 mesh scenario with the 3 mesh scenario at the first (a) and second (b, c) measurement interfaces, placed at 0.40 m away from the
boundary interfaces.

nel height as recorded in the location of the boundary interfaces for
both the single mesh and the multiple mesh. Because the first
boundary interface was positioned upstream of the fire, the tem-
perature information was not analyzed for this position. The veloc-
ity profile at the first interface and the temperature profile at the
second interface show that the boundary interfaces induce negligi-
ble differences from the single mesh model. The velocity at the sec-
ond interface has the same profile and the values differ with a
maximum of 0.2 m/s. As a result, it was deemed appropriate to
use multiple processes in the feasibility calculations, and the grid
was divided further, resulting in 17 meshes.
2.2. 1D model and multiscale coupling
The Heating, Ventilation and Air Conditioning (HVAC) module
has been added to FDS in order to provide coupling between the
solutions of the conservation equations for inlets and outlets of a
HVAC network and the computational grid (Floyd et al., 2013). It
is a 1D model consisting of a network of ducts and nodes, where
the former represents any continuous flow path without node
interruption and the latter represents the point where the duct
joins the FDS computational domain, the ambient or where multi-
ple ducts intersect, and thus form a junction (Floyd, 2011). This
model uses an explicit solver for the conservation equations of
mass and energy along with an implicit solver for the conservation
equation of momentum. There is a caveat of the boundary condi-
tions of the ducts, as no thermal loss takes place through the ducts.
However, as this mainly influences the chimney effect in inclined
tunnels, the impact on the no-slope tunnel used herein is
negligible.
In order to simulate the rest of the 1200 m long tunnel, the duct
and node network was defined as follows. On each side of the FDS
grid, a node was placed at the interface with the FDS grid and the
other one at 400 m distance. The duct connecting them has an area
equal to the cross-sectional area of the tunnel, thus 52 m2, and a
friction coefficient of 0.026, as recommended by Jang and Chen
(2002). This yields a relative roughness of 0.003 according the
Moody chart (for tunnels with wind speeds between 1 and
6.5 m/s). From the relative roughness, the absolute roughness,
which is the input parameter, is obtained and has a value of
k = 0.024 m.
The ventilation system considered in the fire scenarios consists
of 5 pairs of fans at the upstream portal. The pairs of jet fans on the
south side of the tunnel are considered not to be operational. This
ventilation scenario is used for comparison with scenario 2 from
Colella et al. (2010). In order to specify the equivalent flow of 5
pairs of fans into an equivalent fan defined in the duct, a value of
mass flow rate of 300 kg/s was taken from the full-scale CFD
results for scenario 2 (Colella et al., 2010). This value was obtained
in the reference cases built for the validation of multiscale mod-
elling (Colella et al., 2010), cases that serve as a reference point
for this work. For the sake of calculation simplicity, the fan was
defined as having a constant volume flow. However, this way of
imposing a flow condition does not capture the throttling effect
of the fire on the ventilation velocity and it does not take into
account the surrounding pressure field (Floyd et al., 2013). The
throttling effect is the tendency of the fire to resist the airflow
(Vaitkevicius et al., 2015), which means that the same ventilation
scenario cannot be used for fire sizes ranging from small to large
and without causing differences in the flow response. However,
considering that the aim of this study is to assess the feasibility
of implementing multiscale modelling in FDS v.6, not to design a
viable ventilation scenario for a tunnel, using a constant flow is
deemed acceptable.
The information exchange between the two models is illus-
trated in Fig. 5. The initial step is to update the density solution,
which is followed by the calculation of the average values (temper-
ature, pressure, species) at each HVAC node coupled to the main
150 I. Vermesi et al. / Tunnelling and Underground Space Technology 64 (2017) 146–153
Fig. 5. Solving procedure for HVAC + FDS model.
grid gas phase solution. The next step is to perform the equations
for the HVAC solvers, then use this solution to update the boundary
conditions. By updating the divergence, the cycle continues and the
calculation moves on to the next time step.
3. Results and discussion
3.1. Multiscale modelling feasibility in FDS v.6.0
As temperatures and velocities are the most important param-
eters, it is of interest to see how they compare to the results
obtained in the reference work. However, it is important to estab-
lish a few of the differences between the two models that arise
from the different modelling techniques.
The FDS model was designed to investigate fire flows (Floyd
et al., 2013) and has a dedicated combustion model. This simulates
the fire behaviour differently from the volumetric heat source used
to model the fire in Fluent, which is a general CFD tool that can be
used for a variety of situations. The behaviour of the fire in FDS is
influenced by the buoyancy forces as well as by the rate at which
the fire is ’fed’ oxygen by the ventilation system.
Furthermore, there is a difference in the way the two programs
handle meshes. Fluent uses an unstructured grid to be able to sim-
ulate more realistic geometrical shapes, whereas FDS uses a struc-
tured, rectangular grid, and, as such, simplifies the geometry to be
able to perform the calculation on rectangular cells. As Colella et al.
(2010) used Reynolds-Average Navier-Stokes equations, they could
use a finer mesh in Fluent without requiring a computational effort
as great as FDS would require to solve the same mesh. As FDS uses
the large eddy simulation method, it does not average the flow,
thus requiring more time for calculations. Also, Colella et al.
(2010) used a steady state simulation, while the simulation in
FDS is transient and captures the fire for 10 min. Thus, it is obvious
that the results, although averaged, will have some discrepancies.
Fig. 6. Longitudinal slice of the downstream flow behaviour through the tunnel centre line: average temperature in °C; the fire is represented by the white box on the left of
the figure; cross-sectional slices at 10 and 100 m which present the temperature values are shown in Fig. 8.
Taking into account these differences, the longitudinal flow
conditions in the tunnel are presented in Figs. 6 and 7. It is clear
that the flow becomes fully-developed with only small differences
in temperature and velocity across the cross-section as it
approaches the interface with the 1D model. The temperature
and velocity results were evaluated at 10 and 100 m from the cen-
tre of the fire. As seen in Figs. 8 and 9, the flow 10 m downstream of
the fire presents a clear division between the area of the plume and
the rest of the tunnel. The area around the plume has a much
higher velocity than the plume itself, due to the ventilation air
bypassing the plume on its sides. Further away from the fire, at
100 m, the flow becomes more homogeneous, with a higher overall
velocity. This confirms the assumption that the flow develops
enough to be modelled as a one-dimensional flow. The centreline
time-averaged results for a section of the 3D domain that compare
the FDS and ANSYS models are shown in Figs. 10 and 11. The shape
of the temperature distribution has similar characteristics and the
values are very similar, thus proving that multiscale modelling
works as well for FDS as it did for ANSYS Fluent. However, it must
be noted that the velocity in this scenario is much higher than the
reference values because of the way the flow is fixed. If a quadratic
fan curve were to be implemented, the throttling effect of the fire
would interact with the fan flow and the values of the velocity
would be reduced.
3.2. Reduction in computational resources
The most significant advantage of multiscale modelling is
reducing the duration of simulations. As multiple processor calcu-
lations have the same purpose, it is interesting to see which of
these two methods is more efficient. To make a comparison, four
simulations were made: one with the full tunnel using a single
mesh, one with the full tunnel using multiple meshes, one with
multiscale modelling with a single mesh and one with multiscale
modelling combined with multiple meshes. The model was the
same as the one used in the feasibility analysis. The jet fans, which
were implemented as a fixed flow in the multiscale model, were
modelled as simple vents with a volume flow.
For a comparison regarding simulation times, it is important to
specify the computational resources that were available. The sim-
ulations were performed on a single computer with 2.20 GHz CPU
clock and 18 cores. Because of the centrally placed fire, the model
had to be divided into an odd number of meshes. Therefore, 17
meshes were used for the multiple meshes simulations.
Table 2 shows the runtimes of each analyzed model. Using a
single mesh and modelling the entire tunnel length of 1200 m
resulted in a simulation that required an estimated 3 weeks for
completion, which is not practical when having to investigate
many scenarios. The estimation was calculated using the total
required wall clock time as the ratio between the elapsed wall
clock time and elapsed FDS simulation time multiplied by the total
FDS simulation time (Floyd et al., 2013). Trying to overcome this
issue by dividing the grid into multiple meshes halved the runtime,
but it is still impractically long. However, using a multiscale model
reduced the runtime to less than a day, which makes it viable for
design purposes. If multiple processors are available, combining
 I. Vermesi et al. / Tunnelling and Underground Space Technology 64 (2017) 146–153 151

Fig. 7. Longitudinal slice of the downstream flow behaviour through the tunnel centre line: average u-velocity in m/s; the fire is represented by the white box on the left of
the figure; cross-sectional slices at 10 and 100 m which present the temperature values are shown in Fig. 9.

(a) 10 m (b) 100 m

Fig. 8. Cross-sectional slice of the average temperature (in °C) at various distances downstream from the fire, as shown in Fig. 6.

(a) 10 m (b) 100 m

Fig. 9. Cross-sectional slice of the average u-velocity (in m/s) at various distances downstream from the fire, as shown in Fig. 7.

multiple meshes and multiscale modelling results in the most
time-saving solution. The drawback of multiple processor calcula-
tion is obviously the dependence on the available number of pro-
cessors, a problem which does not appear when using multiscale
modelling. This method reduced the simulation time from weeks
to less than a day and can thus be applied on a variety of computer
configurations.
3.3. Validation with experiments and oscillatory mass flows
In order to validate the multiscale model with experiments, a
comparison of cold-flow conditions between the simulation of
the tunnel and experimental measurements was carried out and
discussed by Ang et al. (2016). That work compares measurements
of mass flow through the Dartford tunnel with a full CFD model
and a multiscale model done in FDS v.6.1. It is shown that average
velocities are very similar in the full CFD and multiscale models
and that they compare well with the measurements. The model
presented in the current work formulates the concept for the
model used by the authors of the validation (Ang et al., 2016).
Therefore, the validation is applied for the multiscale model
methodology using FDS v.6.0. However, there were no experimen-
tal measurements of an actual fire in the Dartford tunnel to serve
as validation, so the validation in the cold flow scenario was used.
As shown by Ang (2014), when adding a fire load to the multi-
scale model in FDS v.6.1, the mass flows show an oscillating beha-
viour. This can be observed in the results of the current work as
well, which was done in FDS v.6.0, and are worthy of consideration
because they signify a numerical issue within the solver that might
influence the results.
As shown in Fig. 3, this oscillatory mass flow influences both
centreline temperatures and velocities for meshes with medium
152 I. Vermesi et al. / Tunnelling and Underground Space Technology 64 (2017) 146–153
Fig. 10. Temperature slice at 600 s - 5 working jet fan pairs, 30 MW fire, 3D domain
of 400 m, cell size of 0.4 m, time-averaged results; the length of the FDS domain is
identical to the one in Fluent, with the difference arising from image scaling.
Fig. 11. U-velocity (x-velocity) slice at 600 s - 5 working jet fan pairs, 30 MW fire,
3D domain of 400 m, cell size of 0.4 m, time-averaged results; the length of the FDS
domain is identical to the one in Fluent, with the difference arising from image
scaling.
Table 2
Comparison between the runtimes using multiscale modelling and multiple processor
calculation.
Tunnel model Runtime [h]
Full tunnel: single mesh 519 (est.)
Full tunnel: multiple (17) mesh 280 (est.)
Multiscale model (3D length of 400 m): single mesh 13.2
Multiscale model (3D length of 400 m): multiple (17) mesh 6.1
or fine cell size, but are not appearing in the mesh with large cell
size (0.8 m). Instead of reaching steady-state, the mass flow rate
varies in time following a periodic oscillations. This is a numerical
issue encountered in the solver of both FDS v.6.0 and 6.1 and has to
be investigated in the newer versions of FDS, which have been
released since the publication of this work.
The oscillatory mass flow is an issue of multiscale modelling in
FDS v.6.0 and 6.1, regardless of the way the ventilation flow is
modelled. This work uses a simple constant mass flow as a bound-
ary at the 1D-3D interface, whereas Ang (Ang, 2014) uses a 3D
domain to model the flow coming from the jet fans. This issue is
specific to the multiscale model.
4. Conclusions
Based upon the current feasibility analysis, multiscale mod-
elling is suitable for FDS v.6.0 using the HVAC component. The
temperature results show good agreement with the reference work
values, displaying a difference of just 2%. The differences in the
velocity results arise from the difference in flow definition between
the two models (FDS and ANSYS Fluent). The fixed flow used herein
(FDS) induces a much larger flow velocity than the quadratic flow
used in the reference work (ANSYS Fluent), and as a result the
velocity is much higher in the current case.
When implemented in FDS v.6.0, multiscale modelling provides
the fastest way to simulate a large number of fire scenarios. It does
not depend on the computational resources as much as the multi-
ple processor calculation does. It also reduces the simulation time
significantly while keeping the precision of the results. Compared
to the original full-scale tunnel with a single mesh, multiscale
modelling using a single mesh reduced the simulation duration
by 97%, so a simulation took a few hours instead of weeks. As a
comparison, a multiple processor calculation where the model
was divided into 17 meshes only resulted in a 46% reduction in
computational time. If possible, the most efficient way is to com-
bine both methods, and in the current study a 99% reduction in
the necessary computational time was obtained for this combina-
tion. Of course, these values are subject to the available computing
resources. However, the ranking between them is independent of
the computational power.
It is important to specify that the purpose of this work was to
assess the feasibility of multiscale modelling in FDS v.6.0 and com-
pare its simulation time reduction to the one given by using mul-
tiple processor calculation. The study was not aimed at analyzing
the efficiency of a particular fire safety design strategy. For such
considerations, flow conditions have to be improved. For the future
work, defining a quadratic flow using a fan curve is extremely
important in order to capitalize on the interaction between the
3D model and the 1D HVAC network. In addition, the oscillatory
mass flows that result from a numerical issue have to be checked
in the latest version of FDS.
Keeping in mind the oscillations and the lack of validation in
fire conditions, this method can be cautiously adopted to look at
steady state conditions near the fire. The information can be used
in the design of the tunnel structure, as well as in establishing the
passive fire protection (PFP) measures to be used in the fire safety
design of the tunnel. The heat losses can be accurately modelled in
the 3D domain of the model. Also, the 3D domain is large enough
to contain the smoke, thus the potential issue of defining the walls
of the tunnel as adiabatic is acceptable, since it provides a more
conservative estimates.
Acknowledgements
The authors would like to thank NIST and VTT for providing the
FDS v.6.0 code as open-source.
References
Ang, C., 2014. Investigation of a Computationally Efficient Multi-Scale Modelling
Method in Long Tunnels for Fire Dynamics Simulator 6. Master’s Thesis,
Imperial College London.
Ang, C., Rein, G., Peiro, J., Harrison, R., 2016. Simulating longitudinal ventilation
flows in long tunnels: comparison of full CFD and multi-scale modelling
approaches in FDS6. Tunn. Undergr. Space Technol. 52, 119–126.
Carvel, R., Beard, A., 2005. The Handbook of Tunnel Fire Safety. Thomas Telford
Publishing.
Colella, F., Rein, G., Borchiellini, R., Torero, J.L., 2010. A novel multiscale
methodology for simulating tunnel ventilation flows during fires. Fire
Technol. 47 (1), 221–253.
 I. Vermesi et al. / Tunnelling and Underground Space Technology 64 (2017) 146–153
Colella, F., Rein, G., Carvel, R., Reszka, P., Torero, J.L., 2010. Analysis of the ventilation
systems in the Dartford tunnels using a multi-scale modelling approach. Tunn.
Undergr. Space Technol. 25, 423–432.
Colella, F., Rein, G., Verda, V., Borchiellini, R., 2011. Multiscale modeling of transient
flows from fire and ventilation in long tunnels. Comput. Fluids 51 (1), 16–29.
Colella, F., Rein, G., Borchiellini, R., Verda, V., 2012. One-dimensional and multi-
scale modelling of tunnel ventilation and fires. In: Handbook of Tunnel Fire
Safety, pp. 365–390.
Floyd, J., 2011. Coupling a network HVAC model to a computational fluid dynamics
model using large eddy simulation. Fire Safety Sci. (10), 459–470
Floyd, J., Forney, G., Hostikka, S., Korhonen, T., McDermott, R., McGrattan, K.,
Weinschenk, C., 2013. NIST Special Publication 1019. Fire Dynamics Simulator,
User’s Guide, sixth ed..
Floyd, J., Forney, G., Hostikka, S., Korhonen, T., McDermott, R., McGrattan, K.,
Weinschenk, C., 2013. NIST Special Publication 1019. Fire Dynamics Simulator
Technical Reference Guide. Mathematical Model, sixth ed., vol. 1.
Jain, S., Kumar, S., Kumar, S., Sharma, T., 2008. Numerical simulation of fire in a
tunnel: comparative study of CFAST and CFX predictions. Tunn. Undergr. Space
Technol. 23, 160–170.
153
Jang, H.-M., Chen, F., 2002. On the determination of the aerodynamic coefficients of
highway tunnels. J. Wind Eng. Ind. Aerodyn. 90 (8), 869–896.
Migoya, E., Crespo, A., Garcia, J., Hernandez, J., 2009. A simplified model of fires in
road tunnels. Comparison with three-dimensional models and full-scale
measurements. Tunn. Undergr. Space Technol. 24, 37–52.
Petterson, N., 2002. Assessing the Feasability of Reducing the Grid Resolution in FDS
Field Modeling. Tech. Rep., University of Canterbury.
Vaitkevicius, A., Colella, F., Carvel, R., 2015. Investigating the throttling effect in
tunnel fires. Fire Technol., 1–10
Vega, M., Diaz, K., Oro, J., Tajadura, R., Morros, C., 2008. Numerical 3D simulation of
a longitudinal ventilation system: memorial tunnel case. Tunn. Undergr. Space
Technol. 23, 539–551.
Versteeg, H., Malalasekera, W., 2007. An Introduction to Computational Fluid
Dynamics: The Finite Volume Method. Prentice Hall.
Weisenpacher, P., Halada, L., Glasa, J., 2011. Computer Simulation of Fire in a Tunnel
Using Parallel Version of FDS. In: Seventh Mediterranean Combustion
Symposium, Cagliari.